This paper, by Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel of OpenAI and UC Berkeley, was submitted to arXiv on March 20, 2017 and presented at IROS 2017. It introduced and popularized domain randomization, a now-standard technique for bridging the reality gap - the persistent mismatch between cheap, fast physics simulators and the messy real world that causes models trained in simulation to fail on real robots.
The core insight is counterintuitive. Rather than making the simulator more photorealistic to match reality, you make it wildly varied: randomize textures, colors, lighting, camera position, and object placement across an enormous range, including non-realistic random textures. With enough variability in training, the real world looks to the model like just one more variation it has already seen, so no special adaptation is needed at deployment.
The team demonstrated this on object localization, a building block for robotic grasping. Using only data from a simulator with random textures and no real images at all, they trained a detector that localized objects to within 1.5 centimeters in the real world and stayed robust to distractors and partial occlusions. Domain randomization went on to underpin OpenAI’s later Rubik’s Cube robot hand work and became a default tool whenever robots are trained mostly in simulation.